eazyware
Engineering·June 19, 2023·11 min read

AI recommendation systems: 2026 patterns

Content-based, collaborative, hybrid, neural. Recommendation systems have been ML for a decade; AI era patterns and what changed.

KR
Kushal R.
Engineering lead

Recommendation systems have been machine learning for over a decade. The AI era added better embeddings, sequence models, and conversational recommendation. Production systems combine multiple approaches. This post is the patterns that dominate production recommendation in 2026 and what the LLM era specifically changed.

Approaches
Recommendation approaches Content-based Item features match Works for new items Cold-start friendly Collaborative User similarity Matrix factorization Needs interaction data Hybrid Combine approaches Ensemble ranking Industry standard Neural / AI-era Two-tower, transformers Sequence models LLM-assisted What LLM era added Better item embeddings from LLM encoders; richer semantic matches Explanation and justification in natural language for recs Conversational discovery — recs as part of dialog, not just ranked list
Content-based: item features. Collaborative: user similarity. Hybrid: combined. Neural / AI-era: two-tower, transformers, LLM-assisted.

Content-based filtering

Recommend items similar to what user interacted with. 'If you liked X, you'll like Y because Y has similar features.'

Works for new items. Features known even without interaction data.

Cold-start friendly. Works for items with no interaction history yet.

Limitations. Restricted to items with known features. Doesn't capture taste beyond explicit features.

Collaborative filtering

Recommend items other similar users liked. 'Users who liked X also liked Y.'

Matrix factorization. Classic collaborative filtering. User-item interaction matrix decomposed.

Needs interaction data. Cold-start problem for new users and new items.

Scalability. Works at scale with right infrastructure. Spotify, Netflix, YouTube use variants.

Hybrid approaches

Combine content and collaborative. Mitigates cold-start; captures both explicit features and emergent patterns.

Industry standard. Most production recommenders are hybrid.

Ensemble ranking. Multiple models produce scores; meta-model combines.

Neural / AI-era approaches

Two-tower models. User tower and item tower produce embeddings; similarity drives ranking. Scalable for large catalogs.

Transformers. Sequence-aware recommendations. User's interaction history as sequence; predicts next interaction.

Graph neural networks. User-item interactions as graph; GNN propagates preferences. Captures indirect relationships.

LLM-assisted. LLM enrichment of items, generation of natural language explanations, conversational recommendations.

What LLM era specifically added

Better item embeddings. LLM encoders produce richer representations than older embeddings.

Richer semantic matches. 'Like that but more adventurous' — LLM can understand and match on this.

Explanations. Natural language justification for recommendations. 'I recommend X because you liked Y and similar themes of Z.'

Conversational discovery. Chat interfaces replacing pure ranked lists for some use cases. User iterates preferences.

Cold-start for new items. LLM can extract item features from text, images, reviews. Bootstraps new items into recommendation system.

Production architecture

Candidate generation. Fast retrieval of O(hundreds) candidates from O(millions) items. Two-tower or similar.

Ranking. Deep ranking models score candidates considering context, user history, time of day, device. Typically gradient boosting, transformers, or MoE.

Reranking. Business rules, diversity, exploration — applied after main ranking.

Explainability. Why this recommendation? Increasingly important for trust.

Challenges

Cold-start. New users, new items. Content-based and LLM enrichment help.

Feedback loops. Recommendations affect what users see; users click what they see; creates feedback. Exploration mechanisms counter.

Filter bubbles. Personalization can narrow exposure. Diversity requirements balance.

Privacy. Recommendation systems use data; privacy-preserving techniques (federated learning, on-device models) emerging.

Fairness. Recommendations may disadvantage some items or creators. Active research area.

Evaluation

Offline metrics. Precision@K, Recall@K, NDCG, MAP. Predict user engagement.

Online metrics. Click-through rate, engagement, revenue, retention. Ultimately what matters.

A/B testing. Standard for recommendation changes. See A/B testing post.

Counterfactual evaluation. Estimating what would have happened with different recommendations. Increasingly important.

Industry examples

Netflix. Deep personalization; conversational recommendations emerging.

Spotify. Music recommendations; now conversational via AI DJ feature.

Amazon. Product recommendations; LLM-powered shopping assistant emerging.

TikTok. Short video recommendations — arguably most sophisticated in production.

Enterprise uses. Customer recommendation engines for e-commerce; content recommendations for news/media; product recommendations for B2B.

Read next
AI for search ranking: retrieval + rerank patterns
Read next
AI usage analytics: what to measure, how to act on it
Tags
recommendationsMLpersonalization
/ Next step

Want to talk about this?

We love debating this stuff. 30-minute call, no pitch, just engineering conversation.

~4h
avg response
Q2 '26
next slot
100%
NDA on request